AI Providers, Streaming & DOCX export

AI / Server

Providers, streaming & DOCX export

Server-side companions to the AI Toolkit, all shipped in the RichTextBox.AspNetCore NuGet package (1.0.0-preview.11+). One-line DI to wire a real provider, SSE streaming end-to-end, and an HTML → DOCX endpoint.

Provider resolvers Streaming demo DOCX demo

Providers Streaming DOCX export Custom resolvers

Built-in provider resolvers

Three drop-in IRichTextBoxAiResolver implementations using HttpClientFactory, cancellation, and ILogger<T>. Pick one and register in Program.cs:

using RichTextBox.AiResolvers;

builder.Services.AddRichTextBox();
builder.Services.AddRichTextBoxOpenAiResolver(opts =>
{
    opts.ApiKey = builder.Configuration["OpenAI:ApiKey"];
    opts.Model  = "gpt-4o-mini";
});

Swap AddRichTextBoxOpenAiResolver for AddRichTextBoxAnthropicResolver or AddRichTextBoxAzureOpenAiResolver. All three share AiResolverOptions:

public class AiResolverOptions
{
    public string  ApiKey        { get; set; }       // required
    public string? Model         { get; set; }       // provider-specific default
    public string? BaseUrl       { get; set; }       // for proxies / Ollama / Groq
    public int     MaxTokens     { get; set; } = 1024;
    public double  Temperature   { get; set; } = 0.4;
    public string? SystemSuffix  { get; set; }       // appended to every mode's system prompt
    public TimeSpan Timeout      { get; set; } = 60s;
}

Azure has three extras on AzureOpenAiResolverOptions: Endpoint (https://my-resource.openai.azure.com), DeploymentName (not model), ApiVersion (optional).

The OpenAiResolver works with any OpenAI-compatible endpoint — override BaseUrl for Groq, Together.ai, or local Ollama.

Streaming responses (Server-Sent Events)

MapRichTextBoxUploads() also maps POST /richtextbox/ai/stream (configurable via RichTextBoxOptions.AiStreamEndpoint). OpenAI and Anthropic resolvers implement IStreamingRichTextBoxAiResolver; non-streaming providers fall back to a single data frame transparently.

Wire format

data: <json-encoded delta string>\n\n
data: <json-encoded delta string>\n\n
...
event: response
data: { ...RichTextBoxAiResponse payload... }\n\n
event: done
data: {}\n\n

Client

var stream = editor.aiToolkit.streamRequest({
    url: "/richtextbox/ai/stream",
    body: { mode: "rewrite", selectionText: "..." },
    onDelta:    function (chunk, full) { /* render progressively */ },
    onResponse: function (payload)     { /* final structured result */ },
    onDone:     function (fullText)    { },
    onError:    function (err)         { }
});
// stream.abort()  â€” cancel mid-response
// stream.promise  â€” resolves with the concatenated full text

The helper uses fetch + ReadableStream. Browsers without streaming fall back to XHR against the non-streaming endpoint automatically.

DOCX export

MapRichTextBoxUploads() also maps POST /richtextbox/export/docx (configurable via DocxExportEndpoint). Built on MIT-licensed DocumentFormat.OpenXml. No external service.

Request

POST /richtextbox/export/docx
Content-Type: application/json

{
    "html": "<h1>Memo</h1><p>Body.</p>",
    "title": "Memo",
    "fileName": "memo.docx"
}

Response

Binary stream with Content-Type: application/vnd.openxmlformats-officedocument.wordprocessingml.document and Content-Disposition: attachment; filename=memo.docx.

Supported HTML

Paragraphs, h1–h6, blockquote, pre, hr
Ordered & unordered nested lists
Tables (with header rows)
Inline: b/strong, i/em, u, s/del, sub, sup, code, br
Hyperlinks, inline images (data URIs only — external URLs skipped for safety)
Style attributes: font-weight, font-style, text-decoration, color, background-color, text-align

Client helper

editor.aiToolkit.exportDocx({
    url:      "/richtextbox/export/docx",
    fileName: "report.docx",
    title:    "Quarterly report"
});

Custom resolvers

For Bedrock, Gemini, your own model, or anything else — implement IRichTextBoxAiResolver (and optionally IStreamingRichTextBoxAiResolver) and register your class. The rest of the AI Toolkit surface is provider-agnostic.

public sealed class MyBedrockResolver : IRichTextBoxAiResolver
{
    public async ValueTask<RichTextBoxAiResponse> ResolveAsync(
        RichTextBoxAiRequest request, CancellationToken cancellationToken = default)
    {
        // ...call Bedrock, map response to RichTextBoxAiResponseBuilder...
    }
}

builder.Services.AddRichTextBox();
builder.Services.AddSingleton<IRichTextBoxAiResolver, MyBedrockResolver>();

Documentation